ensure we don't raise in any `*.native_to_narwhals_dtype`

[dangotbanned]

Related

• fix(DONT MERGE): duckdb>=1.4.1 typing & warnings #3189 (comment)
• Question: How are non-narwhals-supported datatypes handled? akmalsoliev/Validoopsie#30

Description

(#3189 (comment))

an intentional choice to be permissive-in & strict-out

I think the idea is:

• permissive-in: if a user does nothing with a column (i.e. just passes it around), it makes not sense for us to raise an exception early on
• strict-out: we don't really have any other options - if a backend does not support a dtype, what else should we do if not raising an error?

We should add tests so that regressing on this (like I tried to) is caught locally.

The intended behavior is we return nw.Unknown()

[MarcoGorelli]

on the pandas side we have

narwhals/tests/frame/schema_test.py

Lines 197 to 199 in 5838dde

	def test_unknown_dtype() -> None:
	df = pd.DataFrame({"a": pd.period_range("2000", periods=3, freq="M")})
	assert nw.from_native(df).schema == {"a": nw.Unknown}

[dangotbanned]

#3197 (comment)
Yeah I'm thinking something like that, but for each backend and probably more than 1 dtype 😄

[dangotbanned]

I just spent a while trying to do a repro for the ibis.expr.datatypes.Interval error in (#3189 (comment))

But I can't seem to do it 🤔

import ibis
import polars as pl
from ibis.common.temporal import IntervalUnit
from ibis.expr.datatypes import Interval

import narwhals as nw

df_ibis: ibis.Table = (
    nw.from_native(pl.select(a=pl.duration(weeks=1))).lazy("ibis").to_native()
)
df_ibis_bad = df_ibis.cast({"a": Interval(unit=IntervalUnit.DAY)})
nw.from_native(df_ibis_bad).collect("polars")


ConversionException: Conversion Error: Unimplemented type for cast (INTERVAL -> INTEGER) when casting from source column a

LINE 1: SELECT to_days(CAST(t0.a AS INTEGER)) AS a FROM ibis_polars_memtable_celwt...

I was thinking we might have been rejecting something similar to pandas.IntervalDtype or pandas.PeriodDtype